AI is accelerating faster than ever, but voters are pushing back. This week we look at Anthropic’s models improving at twice their historical rate, with AI responsible for 80% of it, how $7.5 billion in Nvidia chips reportedly reached China through Jensen Huang’s affair, Claude dreaming to fix its memory, and why US midterm results could slow the data-center build out.
The trend:Anthropic gave METR its internal R&D data, and they concluded that Anthropic is improving its models’ capabilities at twice the historical rate, and that 80% of the acceleration is from AI doing R&D.
The details: Between August and the end of October, staff from the independent evaluator spent 11 weeks assessing the extent to which AI is accelerating AI R&D, as significant acceleration could lead to a “sudden emergence of extreme and risky capabilities.” Benchmark scores for Anthropic’s models have been climbing at 1.9x the pre-October 2025 rate, with METR confident that this wasn’t driven by an acceleration in the scaling of training compute. They were careful to say that AI isn’t making fully autonomous contributions to model training yet, but researchers have described “signs of life” that it could soon.
Why it matters: METR also found no obvious bottlenecks that would slow acceleration in the next 12 months, meaning there could be a significant risk of faster acceleration. Anthropic also redacted a majority of the evidence that METR found most informative, and canceled a planned audit on what percent of compute was allocated to capabilities vs safety halfway through.
The details: During the relationship, Alice Huang’s efforts got roughly 240,000 chips into China at a time when orders were affected by the export curbs. American officials received a briefing saying that she worked for China’s Ministry of State Security, and that Nvidia’s CEO had been the target of a foreign intelligence operation. President Trump said on Monday that Jensen Huang told him he “had no idea that she was a spy.”
Why it matters: Nvidia is already opening an investigation into its own CEO, the FBI is also looking into the affair, and there’s now a proposed class action lawsuit accusing Nvidia and its CEO of securities fraud, after the stock already lost $190 billion in a single day.
Anthropic has shared how Claude agents are using idle time between interactions to fix and reorganize their memory, a process which they call dreaming. When dreaming, Claude checks notes from previous sessions against original transcripts to catch mistakes. Anthropic says that the hardest part of memory is that it’s a prediction about what future sessions might need, but newer models have gotten much better at judging what events to log. Testing dreaming on Pokémon, Fable 5 earned four gym badges with only 10 memory files, while Sonnet 3.7 only got one badge using 413 memory files. Telling a model what memory is for, but not what to do with it, worked far better than rules like “record only failures,” which made stronger models worse.
As we talked about last week, the US midterms doubled as a referendum on data centers, and the results were overwhelmingly against the AI build-out. Texas, Georgia, and Ohio (which account for over 40% of planned US data center capacity) were won by Democrat candidates who had been critical of construction and pledged to slow or pause their rollout. Dozens of anti data-center ballot measures also passed, signaling growing bipartisan opposition. Sequoia Capital partner Pat Grady said that “this is going to probably elongate the timelines for the data-center build outs” and OpenAI leadership was reportedly discussing whether to reset next year’s capacity expectations on Wednesday morning.
%201.svg)
